US11301214B2 - Device for performing multiply/accumulate operations - Google Patents

Device for performing multiply/accumulate operations Download PDF

Info

Publication number
US11301214B2
US11301214B2 US16/897,190 US202016897190A US11301214B2 US 11301214 B2 US11301214 B2 US 11301214B2 US 202016897190 A US202016897190 A US 202016897190A US 11301214 B2 US11301214 B2 US 11301214B2
Authority
US
United States
Prior art keywords
input
arguments
arg2
arg1
pairs
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
US16/897,190
Other languages
English (en)
Other versions
US20210382689A1 (en
Inventor
Mankit Lo
Meng Yue
Jin Zhang
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
VeriSilicon Microelectronics Shanghai Co Ltd
Verisilicon Holdings Co Ltd Cayman Islands
Verisilicon Holdings Co Ltd USA
Original Assignee
VeriSilicon Microelectronics Shanghai Co Ltd
Verisilicon Holdings Co Ltd Cayman Islands
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by VeriSilicon Microelectronics Shanghai Co Ltd, Verisilicon Holdings Co Ltd Cayman Islands filed Critical VeriSilicon Microelectronics Shanghai Co Ltd
Priority to US16/897,190 priority Critical patent/US11301214B2/en
Assigned to VERISILICON MICROELECTRONICS (SHANGHAI) CO., LTD., VERISILICON HOLDINGS CO., LTD. reassignment VERISILICON MICROELECTRONICS (SHANGHAI) CO., LTD. ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: YUE, Meng, LO, MANKIT, ZHANG, JIN
Priority to JP2021094223A priority patent/JP2021197172A/ja
Priority to KR1020210072852A priority patent/KR20210152956A/ko
Priority to EP21178334.5A priority patent/EP3923132B1/en
Priority to CN202110635294.5A priority patent/CN113778376A/zh
Publication of US20210382689A1 publication Critical patent/US20210382689A1/en
Application granted granted Critical
Publication of US11301214B2 publication Critical patent/US11301214B2/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/52—Multiplying; Dividing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/544—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices for evaluating functions by calculation
    • G06F7/5443—Sum of products
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F5/00—Methods or arrangements for data conversion without changing the order or content of the data handled
    • G06F5/06—Methods or arrangements for data conversion without changing the order or content of the data handled for changing the speed of data flow, i.e. speed regularising or timing, e.g. delay lines, FIFO buffers; over- or underrun control therefor
    • G06F5/065—Partitioned buffers, e.g. allowing multiple independent queues, bidirectional FIFO's
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/50—Adding; Subtracting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/52—Multiplying; Dividing
    • G06F7/523—Multiplying only
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/52—Multiplying; Dividing
    • G06F7/523—Multiplying only
    • G06F7/53—Multiplying only in parallel-parallel fashion, i.e. both operands being entered in parallel
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
    • G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
    • G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
    • G06F7/52—Multiplying; Dividing
    • G06F7/523—Multiplying only
    • G06F7/53—Multiplying only in parallel-parallel fashion, i.e. both operands being entered in parallel
    • G06F7/5324—Multiplying only in parallel-parallel fashion, i.e. both operands being entered in parallel partitioned, i.e. using repetitively a smaller parallel parallel multiplier or using an array of such smaller multipliers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/22—Microcontrol or microprogram arrangements
    • G06F9/28—Enhancement of operational speed, e.g. by using several microcontrol devices operating in parallel
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30003—Arrangements for executing specific machine instructions
    • G06F9/30007—Arrangements for executing specific machine instructions to perform operations on data operands
    • G06F9/30021—Compare instructions, e.g. Greater-Than, Equal-To, MINMAX
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G06F9/5044—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals considering hardware capabilities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5061—Partitioning or combining of resources
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means

Definitions

  • This invention relates to systems and methods for performing high volumes of mathematical operations.
  • GPUs graphics processing units
  • processing pipelines may each be configured to perform a mathematical function.
  • large amounts of data may be processed in parallel.
  • GPUs are also often used for other applications, particularly artificial intelligence.
  • FIG. 1 is a schematic block diagram of a computer system suitable for implementing methods in accordance with embodiments of the invention
  • FIG. 2 is a schematic block diagram of a multiply/accumulate circuit in accordance with an embodiment of the present invention
  • FIG. 3 is a process flow diagram of a method for processing input arguments in the multiply/accumulate circuit in accordance with an embodiment of the present invention.
  • FIG. 4 is a process flow diagram of a method for post-processing products to be accumulated in the multiply/accumulate circuit in accordance with an embodiment of the present invention.
  • Embodiments in accordance with the present invention may be embodied as an apparatus, method, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.), or an embodiment combining software and hardware aspects that may all generally be referred to herein as a “module” or “system.” Furthermore, the present invention may take the form of a computer program product embodied in any tangible medium of expression having computer-usable program code embodied in the medium.
  • a computer-readable medium may include one or more of a portable computer diskette, a hard disk, a random access memory (RAM) device, a read-only memory (ROM) device, an erasable programmable read-only memory (EPROM or Flash memory) device, a portable compact disc read-only memory (CDROM), an optical storage device, and a magnetic storage device.
  • a computer-readable medium may comprise any non-transitory medium that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
  • Computer program code for carrying out operations of the present invention may be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, Smalltalk, C++, or the like and conventional procedural programming languages, such as the “C” programming language or similar programming languages.
  • the program code may execute entirely on a computer system as a stand-alone software package, on a stand-alone hardware unit, partly on a remote computer spaced some distance from the computer, or entirely on a remote computer or server.
  • the remote computer may be connected to the computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).
  • LAN local area network
  • WAN wide area network
  • Internet Service Provider for example, AT&T, MCI, Sprint, EarthLink, MSN, GTE, etc.
  • These computer program instructions may also be stored in a non-transitory computer-readable medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of manufacture including instruction means which implement the function/act specified in the flowchart and/or block diagram block or blocks.
  • the computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
  • FIG. 1 is a block diagram illustrating an example computing device 100 .
  • Computing device 100 may be used to perform various procedures, such as those discussed herein.
  • Computing device 100 can function as a server, a client, or any other computing entity.
  • Computing device can perform various monitoring functions as discussed herein, and can execute one or more application programs, such as the application programs described herein.
  • Computing device 100 can be any of a wide variety of computing devices, such as a desktop computer, a notebook computer, a server computer, a handheld computer, tablet computer and the like.
  • Computing device 100 includes one or more processor(s) 102 , one or more memory device(s) 104 , one or more interface(s) 106 , one or more mass storage device(s) 108 , one or more Input/Output (I/O) device(s) 110 , and a display device 130 all of which are coupled to a bus 112 .
  • Processor(s) 102 include one or more processors or controllers that execute instructions stored in memory device(s) 104 and/or mass storage device(s) 108 .
  • Processor(s) 102 may also include various types of computer-readable media, such as cache memory.
  • Memory device(s) 104 include various computer-readable media, such as volatile memory (e.g., random access memory (RAM) 114 ) and/or nonvolatile memory (e.g., read-only memory (ROM) 116 ). Memory device(s) 104 may also include rewritable ROM, such as Flash memory.
  • volatile memory e.g., random access memory (RAM) 114
  • nonvolatile memory e.g., read-only memory (ROM) 116
  • Memory device(s) 104 may also include rewritable ROM, such as Flash memory.
  • Mass storage device(s) 108 include various computer readable media, such as magnetic tapes, magnetic disks, optical disks, solid-state memory (e.g., Flash memory), and so forth. As shown in FIG. 1 , a particular mass storage device is a hard disk drive 124 . Various drives may also be included in mass storage device(s) 108 to enable reading from and/or writing to the various computer readable media. Mass storage device(s) 108 include removable media 126 and/or non-removable media.
  • I/O device(s) 110 include various devices that allow data and/or other information to be input to or retrieved from computing device 100 .
  • Example I/O device(s) 110 include cursor control devices, keyboards, keypads, microphones, monitors or other display devices, speakers, printers, network interface cards, modems, lenses, CCDs or other image capture devices, and the like.
  • Display device 130 includes any type of device capable of displaying information to one or more users of computing device 100 .
  • Examples of display device 130 include a monitor, display terminal, video projection device, and the like.
  • a graphics-processing unit (GPU) 132 may be coupled to the processor(s) 102 and/or to the display device 130 .
  • the GPU may be operable to render computer generated images and perform other graphical processing.
  • the GPU may include some or all of the functionality of a general-purpose processor, such as the processor(s) 102 .
  • the GPU may also include additional functionality specific to graphics processing.
  • the GPU may include hard-coded and/or hard-wired graphics function related to coordinate transformation, shading, texturing, rasterization, and other functions helpful in rendering a computer generated image.
  • Interface(s) 106 include various interfaces that allow computing device 100 to interact with other systems, devices, or computing environments.
  • Example interface(s) 106 include any number of different network interfaces 120 , such as interfaces to local area networks (LANs), wide area networks (WANs), wireless networks, and the Internet.
  • Other interface(s) include user interface 118 and peripheral device interface 122 .
  • the interface(s) 106 may also include one or more user interface elements 118 .
  • the interface(s) 106 may also include one or more peripheral interfaces such as interfaces for printers, pointing devices (mice, track pad, etc.), keyboards, and the like.
  • Bus 112 allows processor(s) 102 , memory device(s) 104 , interface(s) 106 , mass storage device(s) 108 , and I/O device(s) 110 to communicate with one another, as well as other devices or components coupled to bus 112 .
  • Bus 112 represents one or more of several types of bus structures, such as a system bus, PCI bus, IEEE 1394 bus, USB bus, and so forth.
  • a processor 102 may include a cache 134 , such as one or both of a L1 cache and an L2 cache.
  • a GPU 132 may likewise include a cache 136 that may likewise include one or both of a L1 cache and an L2 cache.
  • programs and other executable program components are shown herein as discrete blocks, although it is understood that such programs and components may reside at various times in different storage components of computing device 100 , and are executed by processor(s) 102 .
  • the systems and procedures described herein can be implemented in hardware, or a combination of hardware, software, and/or firmware.
  • one or more application specific integrated circuits (ASICs) can be programmed to carry out one or more of the systems and procedures described herein.
  • a GPU 132 or other component of the computing device 100 may include the components shown in FIG. 2 .
  • buffers 200 , 202 may store arguments that are to be the subject of a multiply/accumulate operation.
  • one buffer 200 may store a coefficient for implementing a graphics processing operation (e.g., a kernel), artificial intelligence operation (e.g., as part of a convolution neural network).
  • the other buffer 202 may store values (often referred to as the “activation”) to be multiplied by the coefficients. This is, of course, only an example, and any values may be loaded into the buffers 200 , 202 and be the subject of a multiply/accumulate operation.
  • the buffers 200 , 202 may be defined as a portion of memory (RAM 114 ) or portion of a cache 134 or 136 .
  • Each value retrieved from the buffers 200 , 202 may be input to a separator 204 .
  • the separator 204 converts all values into unsigned values.
  • values may be represented in the following format: [type][magnitude].
  • the [type] field indicates whether the bits in [magnitude] represent a signed or unsigned number, e.g. 0 indicating unsigned and 1 indicating signed. Where the [type] field indicates a signed value, the most significant bit (MSB) in the [magnitude] field will be 1 for negative numbers and 0 for positive numbers where 2's compliment representation is used.
  • MSB most significant bit
  • the output of the separator 204 will be a sign 206 and a magnitude 208 for the value from the buffer 200 and a sign 210 and a magnitude 212 for the value from the buffer 202 .
  • the signs 206 , 210 and magnitudes 208 , 212 may then be input to a checker 214 .
  • the checker 214 evaluates the magnitudes 208 , 212 to detect certain cases that require special processing.
  • the number of bits used to represent the magnitudes 208 , 212 may be limited to a number N of bits.
  • N may be the number of bits of values that are actually input to the multiplication circuit and may be less than the number of bits in the [magnitude] field.
  • MaxSign The largest positive number that may be represented by the [magnitude] field of a signed value is referred to herein as MaxSign and may be defined as 2 ⁇ circumflex over ( ) ⁇ N ⁇ 1, where the number of bits in the [magnitude field] is N+1.
  • the checker 214 may detect instances where a magnitude 208 , 212 exceeds MaxSign and make adjustments in response. The manner in which this scenario is handled is described below with respect to FIG. 3 .
  • the maximum value represented with N+1 bits is 2 ⁇ circumflex over ( ) ⁇ (N+1) ⁇ 1. Accordingly, values from 2 ⁇ circumflex over ( ) ⁇ N to 2 ⁇ circumflex over ( ) ⁇ (N+1) ⁇ 1 are not representable with N bits.
  • the checker 214 may likewise discuss when a magnitude 208 , 212 for an unsigned value exceeds MaxSign and make adjustments accordingly, as described below with respect to FIG. 3 .
  • the output of the checker 214 are pairs of arguments, e.g. for a pair of values from the buffers 200 , 202 , the output of the checker 214 is one or more pairs of arguments to be input to a sequencer 216 .
  • the sequencer 216 submits the pairs of arguments to computation units 218 . In particular, there may be multiple computation units, e.g. 8, 64, 1024, or any number of computation units.
  • the sequencer 216 implements logic to submit arguments to a correct computation unit. In particular, the sequencer, 216 ensures that arguments for a pair of values from buffers 200 , 202 are submitted to the computation unit 218 accumulating multiply/add results for that pair of values.
  • each value in the output matrix is a result of a dot product of a row of a first matrix with a column of a second matrix.
  • the sequencer 216 submits arguments for pairs of input values from the buffers 200 , 202 such that each computation unit 218 may accumulate a sum of the products of the elements of a particular row and the elements of a column corresponding to that row.
  • this is only one example and the sequencer 216 can be programmed to accumulate products according to any desired function.
  • Each computation unit 218 may include an N-bit multiplier 220 that takes as inputs a pair of arguments from the sequencer 216 and an adder 222 that takes as inputs a product of the multiplier 220 and the contents of an accumulation buffer 224 .
  • the output of the adder 222 is written back to the accumulation buffer 224 .
  • the results of the accumulation buffer 224 may be read by a controller of a GPU 132 , a CPU 102 according to control of an application, or according to any approach known in the art to retrieving and processing results of a multiply/accumulate operation.
  • the adder 222 may further take as inputs the signs of the input arguments as separated by the separator 204 .
  • the illustrated method 300 may be executed by the checker 214 in order to determine whether to divide an input magnitude 208 , 212 into two arguments or to output a single argument including the input magnitude 208 , 212 .
  • the method 300 may be executed for each input magnitude 208 , 212 , referred to below as “the input magnitude.”
  • the method 300 may include receiving 302 the magnitude and type of the input magnitude from the separator 204 . If the type is found 304 to be signed, the method 300 may include obtaining 306 the absolute value of the input magnitude.
  • the method 300 may then include inputting 312 Arg_1 and Arg_2 to the sequencer 216 .
  • the method 300 may include evaluating 318 whether the input magnitude is larger than MaxSign. If so, then two arguments are set according to steps 318 and 320 : Arg_1 is set equal to the input argument less MaxSign and Arg_2 is set equal to MaxSign.
  • Arg_1 and Arg_2 are then input 322 to the sequencer. If the input magnitude is not found 316 to exceed MaxSign it is input as the argument (Arg) to the sequencer 216 .
  • Inputting the arguments at steps 312 , 314 , 322 , and 324 may be performed in a coordinated manner. Specifically, the argument or arguments as determined for the input magnitude for a value from buffer 200 may be input to the sequencer 216 in coordination with the argument or arguments as determined for the input magnitude for a corresponding value from the buffer 202 .
  • a first value and a second value to be multiplied together may be retrieved from the buffers 200 , 202 , respectively, and be processed by the separator 204 and checker 214 .
  • the pairs of arguments that will be input to the sequencer 216 for various outcomes of the method 300 for the first and second values are described in Table 1.
  • the possible outcomes of the method 300 are either a single output argument designated as Arg1 (step 312 or 324 ) or two output arguments designated here as Arg1_1 and Arg1_2 (step 312 or step 322 ).
  • each pair in parentheses indicates a pair of arguments input to the sequencer 216 that will be multiplied together and accumulated by a computation unit 218 .
  • the sequencer 216 may be programmed to input the pairs of arguments to the same computation unit 218 that corresponds to the first and second values.
  • the sequencer 216 may likewise associate the signs of each argument with it. In particular, where a signed value is split into two arguments Arg_1, Arg_2, the sign will be associated by the sequencer 216 with both of those arguments in all of the argument pairs including either of the two arguments Arg_1, Arg_2.
  • FIG. 4 illustrates a method 400 for performing the multiple/accumulate computation for a pair of arguments input to a computation unit 218 by the sequencer 216 .
  • Each pair of arguments input to the sequencer 216 will be input to the multiplier 220 of one of the computation units, which will then calculate 402 a product P.
  • the method 400 may further include evaluating one or both of the type and sign of the arguments of the pair of arguments. For example, for arguments obtained from unsigned values, the sign at step 404 may be assumed to be positive in all cases. For signed values, the sign will be the sign 206 , 210 as separated from the signed value by the separator 204 .
  • the method 400 may include adjusting 408 the product P. Where there is one negative argument, the sign of P is changed to negative, i.e. P is converted to a negative number, such as according to the 2 's complement definition.
  • the negative product P is then input 410 to the summer 222 , which will then sum the negative product P with the current contents of the accumulation buffer 224 and write the result of the sum to the accumulation buffer 224 .
  • the product P is input 410 to the summer 222 , which will then sum the product P with the current contents of the accumulation buffer 224 and write the result of the sum to the accumulation buffer 224 .
  • multipliers 220 may be made much smaller while still providing the same level of precision using the approach described in FIGS. 2 through 4 .
  • applications such as a GPU, where there are many hundreds or thousands of computation units 218 , this results in a large reduction in circuit area and power usage.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Optimization (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Neurology (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Complex Calculations (AREA)
  • Advance Control (AREA)
  • Image Processing (AREA)
US16/897,190 2020-06-09 2020-06-09 Device for performing multiply/accumulate operations Active US11301214B2 (en)

Priority Applications (5)

Application Number Priority Date Filing Date Title
US16/897,190 US11301214B2 (en) 2020-06-09 2020-06-09 Device for performing multiply/accumulate operations
JP2021094223A JP2021197172A (ja) 2020-06-09 2021-06-04 乗算／累算を実施するためのデバイス
KR1020210072852A KR20210152956A (ko) 2020-06-09 2021-06-04 곱셈/누산 연산을 수행하기 위한 디바이스
EP21178334.5A EP3923132B1 (en) 2020-06-09 2021-06-08 Device for performing multiply/accumulate operations
CN202110635294.5A CN113778376A (zh) 2020-06-09 2021-06-08 改进的用于执行乘法/累加运算的设备

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US16/897,190 US11301214B2 (en) 2020-06-09 2020-06-09 Device for performing multiply/accumulate operations

Publications (2)

Publication Number Publication Date
US20210382689A1 US20210382689A1 (en) 2021-12-09
US11301214B2 true US11301214B2 (en) 2022-04-12

Family

ID=76482988

Family Applications (1)

Application Number Title Priority Date Filing Date
US16/897,190 Active US11301214B2 (en) 2020-06-09 2020-06-09 Device for performing multiply/accumulate operations

Country Status (5)

Country Link
US (1) US11301214B2 (zh)
EP (1) EP3923132B1 (zh)
JP (1) JP2021197172A (zh)
KR (1) KR20210152956A (zh)
CN (1) CN113778376A (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5138570A (en) * 1990-09-20 1992-08-11 At&T Bell Laboratories Multiplier signed and unsigned overflow flags
US6347326B1 (en) * 1999-03-02 2002-02-12 Philips Electronics North America Corporation N bit by M bit multiplication of twos complement numbers using N/2+1 X M/2+1 bit multipliers
EP3374853A1 (en) 2015-11-12 2018-09-19 Arm Ltd Multiplication of first and second operands using redundant representation
US20190392287A1 (en) * 2018-06-22 2019-12-26 Samsung Electronics Co., Ltd. Neural processor
US20200364287A1 (en) * 2019-05-15 2020-11-19 Nxp B.V. Methods and apparatuses involving fast fourier transforms processing of data in a signed magnitude form
US20210073171A1 (en) * 2019-09-10 2021-03-11 Cornami, Inc. Reconfigurable Processor Circuit Architecture

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5138570A (en) * 1990-09-20 1992-08-11 At&T Bell Laboratories Multiplier signed and unsigned overflow flags
US6347326B1 (en) * 1999-03-02 2002-02-12 Philips Electronics North America Corporation N bit by M bit multiplication of twos complement numbers using N/2+1 X M/2+1 bit multipliers
EP3374853A1 (en) 2015-11-12 2018-09-19 Arm Ltd Multiplication of first and second operands using redundant representation
US20190392287A1 (en) * 2018-06-22 2019-12-26 Samsung Electronics Co., Ltd. Neural processor
US20200364287A1 (en) * 2019-05-15 2020-11-19 Nxp B.V. Methods and apparatuses involving fast fourier transforms processing of data in a signed magnitude form
US20210073171A1 (en) * 2019-09-10 2021-03-11 Cornami, Inc. Reconfigurable Processor Circuit Architecture

Also Published As

Publication number Publication date
US20210382689A1 (en) 2021-12-09
EP3923132B1 (en) 2024-01-17
KR20210152956A (ko) 2021-12-16
CN113778376A (zh) 2021-12-10
EP3923132A1 (en) 2021-12-15
JP2021197172A (ja) 2021-12-27

Similar Documents

Publication Publication Date Title
JP6941499B2 (ja) ゼロ係数スキップ畳み込みニューラルネットワークエンジン
CN113330421B (zh) 点积计算器及其操作方法
KR102359265B1 (ko) 프로세싱 장치 및 프로세싱 장치에서 연산을 수행하는 방법
CN108229648B (zh) 匹配存储器中数据位宽的卷积计算方法和装置、设备、介质
US9600236B2 (en) Systems and methods for computing mathematical functions
EP3937010A1 (en) Enhanced multiply accumulate device for neural networks
US11301214B2 (en) Device for performing multiply/accumulate operations
CN110678862A (zh) 用于硬连线乘法器电路的数据打包技术
EP3079056B1 (en) Systems and methods for computing mathematical functions
US9703530B2 (en) Systems and methods for computing mathematical functions
KR102559930B1 (ko) 수학적 함수들을 연산하기 위한 시스템 및 방법들
US9875084B2 (en) Calculating trigonometric functions using a four input dot product circuit
US20210318887A1 (en) Multiplier with Zero Skipping
US8810572B2 (en) Tessellation cache for object rendering
JP2022539554A (ja) 高精度のニューラル処理要素
CN111587441B (zh) 使用以比特值为条件的回归神经网络生成输出示例

Legal Events

Date Code Title Description
AS Assignment

Owner name: VERISILICON HOLDINGS CO., LTD., CAYMAN ISLANDS

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LO, MANKIT;YUE, MENG;ZHANG, JIN;SIGNING DATES FROM 20200604 TO 20200605;REEL/FRAME:052886/0512

Owner name: VERISILICON MICROELECTRONICS (SHANGHAI) CO., LTD., CHINA

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LO, MANKIT;YUE, MENG;ZHANG, JIN;SIGNING DATES FROM 20200604 TO 20200605;REEL/FRAME:052886/0512

FEPP Fee payment procedure

Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY

FEPP Fee payment procedure

Free format text: ENTITY STATUS SET TO SMALL (ORIGINAL EVENT CODE: SMAL); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY

STPP Information on status: patent application and granting procedure in general

Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS

STPP Information on status: patent application and granting procedure in general

Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS

STPP Information on status: patent application and granting procedure in general

Free format text: PUBLICATIONS -- ISSUE FEE PAYMENT VERIFIED

STCF Information on status: patent grant

Free format text: PATENTED CASE